Researchers built a tool to figure out exactly when a human should jump in while an AI agent is mid-task, instead of just waiting or taking over completely.
The study, described in an arXiv paper, ran two rounds of design probes with professional designers working in shared co-creative workspaces, where a person and an agent both edit the same project. In the first round (10 participants), designers spotted chances to contribute by watching the agent work, but held back whenever the agent misread their edits. The researchers then built a second probe tool called CLEO, which tracks user actions and updates its plan accordingly. In the follow-up study (12 participants), the team logged 255 interaction turns and ran recall interviews, sorting the behavior into five action categories, ten specific codes, six triggers, and four enabling factors.
The standout number: designers worked concurrently with the agent, rather than strictly handing off, in 33.73% of turns. That happened when they had a new idea mid-task, wanted to finish a pending subtask themselves, or started an independent version of unfinished work rather than wait.
Most agentic tools still treat collaboration as a relay race: you prompt, the agent runs, you review, repeat. This research suggests that model is wrong for a third of real design work, and that better agents need to track partial user edits and replan on the fly, not just wait for a clean handoff. It is a small, qualitative study of professional designers using research prototypes, not a shipping product, so treat the specific numbers as directional rather than universal to all creative software.